Summary
Fireship covers the story of Ox Alpha, an anonymous frontier-grade model that quietly appeared on OpenRouter and, within six days, served 42 trillion tokens — more than double the next closest model, DeepSeek — before being revealed as GLM 5.3 Flash from Chinese lab Zhipu. The video traces how community sleuthing identified the model’s likely origin: it was the fifth anonymous animal-named model on OpenRouter, and all four predecessors turned out to be Chinese; stack traces matched Zhipu’s API; and a tokenizer fingerprint test returned a match for their GLM series.
The model itself is a natively multimodal mixture-of-experts architecture with 320 billion total parameters, a one-million-token context window, and video input support. It was served entirely on 100,000 Chinese-made chips — a detail that resonated in Silicon Valley given the ongoing chip export debate. On August 26th, Zhipu confirmed the identity and released MIT-licensed weights on Hugging Face. Pricing landed at $0.15 per million input tokens and $0.50 per million output, with a 50% launch discount through September 9th that makes it up to 40 times cheaper than Claude.
Fireship’s hands-on test — modernizing a 2016 AngularJS app called Horse — found the model capable: it detected outdated technologies, fixed a CSS overflow bug using vision, and handled video-to-description tasks well. Weaknesses include slow output speed, verbose narration, and occasional doom loops. The SWE benchmark number that circulated virally (80%) was misleading; the actual score is closer to 58%.
📺 Source: Fireship · Published September 01, 2026
🏷️ Format: News Analysis
